NEURAL NET BASED PRIMING AND MODEL BASED ATR USING MOTION CUES

Object-oriented methodologies such as model-based vision provide a robust and more intelligent solution to the ATR problem. Because these methodologies exploit a priori knowledge of a limited number of target models, while allowing the targets to be oriented towards the viewer in any arbitrary fashion, they result in a powerful ATR system. The central focus of this proposal is on building an automatic target recognition system for identifying mobile targets using an object-image alignment approach. Using this approach, we propose to develop a two stage recognition algorithm. The first stage of this algorithm makes use of the three-point object-image correspondence theorim of Huttenlocher and Ullman (1990) to narrow the search for constrained matching of object-image pairs. The second stage of the algorithm uses the mean-field annealing techique to compute the object-image transformation parameters from multiple matches by minimizing a least-squares measure. For segmentation of the target from its background, we employ a motion-based segmentation algorithm developed earlier by LNK.